After declining by roughly 3% annually between 1920 and 2010, U.S. traffic fatality rates have stagnated—falling significantly behind peer high-income nations like Germany, Canada, and the United Kingdom. As the Labor Day weekend closes out the summer and millions take to the road, we will be reminded that this plateau is an urgent economic and public health crisis, surpassing $1 trillion annually in lost lives, diminished quality of life, medical expenses, and travel disruptions.
Unfortunately, in their search for solutions, policymakers, research funders, and the public are unaware that the academic empirical research used by government agencies to design and justify road safety policy is built on fundamental and fatal statistical flaws.
This research—predominantly published in specialized safety journals by transportation engineers—falls into two main categories: statistical models estimated using data in police accident reports to determine the causes of accidents and controlled crash simulations to determine vehicle safety and the efficacy of safety equipment. Both methodologies suffer from conditioning their analyses on the occurrence of a crash.
When researchers analyze police crash databases, they examine only the motorists who suffered a collision. By definition, this is a non-random sample. This data completely excludes the millions of daily trips where careful drivers avoided collisions and arrived safely. It also excludes trips where risky drivers were under the influence of alcohol or drugs or significantly exceeded the speed limit but did not get into accidents.
These omissions create severe selectivity bias. A driver’s unobserved temperament, judgment, and risk tolerance determine both how they drive and whether they end up in an accident database. When researchers look only at crashed cars to evaluate a tangible safety measure—whether it is expanding highway shoulder widths or mandating driver-assistance features—they cannot separate the physical effectiveness of the reform from the underlying behavior of the driver who may be exposed to it.
Controlled crash tests also fail to account for drivers’ behavior. In a laboratory, a safety device might show a 50% mechanical reduction in fatality risk. But when motorists in the real-world feel protected by passive safety equipment, some respond by driving faster or repeatedly looking at their phones—the well-documented behavioral offset known in economics as the Peltzman effect. A technology that works wonders in a simulation can see its societal benefits largely erased on an actual highway by drivers who use the technology to take more risks.
When a model provides an illusion of mathematical precision while missing the varied effects of real-world behavior, it invites overconfidence. Historically, relying on flawed empirical design parameters for highway policy has led to enormous waste—such as research based on 1950s road tests that miscalculated optimal pavement thickness and saddled taxpayers with decades of inflated maintenance costs. More recent work on seatbelts over-estimated their effectiveness and may have influenced policymakers to prematurely mandate their use.
Scientific progress requires rigorous debate, yet the road safety literature has proven remarkably resistant to examining its foundational assumptions and the implications of its lack of unbiased causal results. A colleague and I pointed out these flaws in the auto safety literature and offered to participate in a debate in one of Elsevier’s automobile safety journals. But the offer spurred no interest among the journals’ editors, who are content to let the matter blow over.
We cannot yet quantify the number of lives that careful causal research might save—unlike the billion-dollar waste caused by pavements that were built too thin—but there is little doubt that safety researchers’ inattention to the flaws we identified contributes to inefficient safety policies. Prematurely mandating expensive automotive equipment or funding costly roadway changes based on inflated benefit estimates impose real costs on households without delivering the promised safety gains.
American motorists have consistently demonstrated a stronger preference for travel volume, speed, and heavy vehicles than their international peers, trading safety for mobility and driving intensity. Because those preferences can offset current safety policies, many of which are not based on reliable research, future breakthrough safety improvements will not come from incremental engineering tweaks or recycled regulatory mandates. Instead, the most promising path lies in the broad adoption of autonomous driving technologies. By shifting the task of driving from risky humans to carefully tested and calibrated software, vehicle autonomy can neutralize risky driving behaviors and expand the safety-mobility frontier in a way public policy alone has failed to achieve.
Until that transition arrives, the first step toward improving safety policy is for federal and state policymakers and research funders to move beyond the statistical illusions of the past by embracing the importance of causal rigor to ensure that public policy can actually save lives.
